AI-assisted engineering with human authority

A replatforming touches a large legacy codebase and many business rules. AI assistants can help, but a generic assistant tends to invent dependencies, ignore conventions and miss behavior hidden in legacy code.

I designed the workflow the team uses on the Flügger migration, with the client's approval to use AI.

Agent rules live next to the code: one file for principles and prohibitions across the monorepo, one per app for its stack and conventions, and a read-only guide to the legacy app. Topic guides for API clients, CMS blocks, domains, state, styling and testing load automatically for the files being edited.

Instead of one generic assistant, agents have narrow jobs: migrating a business domain, checking that legacy behavior still exists in the new app, enforcing code quality and modern APIs, and writing tests.

Any change that spans several files or involves a design decision starts as a short plan (objective, scope, design, test cases, open questions) that is reviewed before implementation. A status table per business domain keeps people and agents from starting work that is already done.

• No new dependencies without explicit approval, and no pipeline changes without instruction.
• No self-merging: AI-generated code goes through the same pull-request review as everything else.
• Pull requests say when AI did a significant part of the work, so reviewers look closely.

I believe the workflow contributed to the new frontend's high unit-test coverage and to consistent code across developers. We have not measured a productivity figure, so I do not quote one.

• Claude
• GitHub Copilot
• Cursor
• MCP
• Next.js 16
• Vitest